Artificial intelligence in higher education: the state of the field

Crompton and Burke update the AI-in-higher-education evidence landscape by reviewing research from 2016 to 2022. The paper is important because it captures the period immediately before and at the edge of the generative AI explosion, showing that higher education AI research was already accelerating before ChatGPT became a public reference point. Using PRISMA principles, the authors identified 138 articles for detailed coding and analysis. The result is a field map that helps readers understand where AI was being studied, who was studying it, and what functions it served.

The review reports rapid growth, especially in 2021 and 2022, with publication numbers rising sharply compared with previous years. It also notes geographical and disciplinary shifts. Earlier reviews had found limited involvement from education departments, but this review reports education as the dominant departmental affiliation among researchers. That makes it a useful companion to Zawacki-Richter et al. 2019, which asked where educators were in AIED research. Crompton and Burke suggest that educator involvement has become more visible, although the field still has gaps.

The paper identifies five major usage categories for AI in higher education: assessment and evaluation, predicting, AI assistant, intelligent tutoring system, and managing student learning. These categories are practical for Research News because they can become tags or filters. They also help AIEDHK separate very different use cases. A predictive model for academic success, a chatbot assistant, an automated assessment tool, and an ITS all require different evidence standards and governance routines.

The review also gives readers a way to see what the field may be over-studying or under-studying. A rapid rise in publications does not automatically mean a mature evidence base. The coding categories reveal where research attention is concentrated and where implementation knowledge may still be thin. For example, if many studies focus on student-facing systems but fewer study instructor workflows, then institutions may lack evidence about teacher workload, professional judgment, and adoption barriers. If language learning is prominent, that may indicate a promising domain, but it should not be generalized to every subject. AIEDHK can use these patterns to plan research coverage that includes gaps, not only popular topics.

The review also highlights who AIED systems are intended to support. Most studies focused on students, with fewer focused on instructors or managers. Language learning was the most common subject domain. Undergraduate students were the most studied group. These findings matter for product strategy: the evidence base may be strongest around student-facing tools and language learning, while teacher-facing and management-facing AI may need more careful development and evaluation. The paper is a current state-of-field entry and should be paired with generative AI work because its search window ends before many ChatGPT-era studies appeared. For AIEDHK, that pairing can show both continuity and rupture in higher education AI research. It also helps readers avoid assuming that today's LLM debate is disconnected from earlier work on assessment, prediction, and tutoring.